diff options
Diffstat (limited to 'core/src/main/scala/spark/scheduler/DAGScheduler.scala')
| -rw-r--r-- | core/src/main/scala/spark/scheduler/DAGScheduler.scala | 412 |
1 files changed, 262 insertions, 150 deletions
diff --git a/core/src/main/scala/spark/scheduler/DAGScheduler.scala b/core/src/main/scala/spark/scheduler/DAGScheduler.scala index 29757b1178..319eef6978 100644 --- a/core/src/main/scala/spark/scheduler/DAGScheduler.scala +++ b/core/src/main/scala/spark/scheduler/DAGScheduler.scala @@ -14,6 +14,7 @@ import spark.partial.ApproximateEvaluator import spark.partial.PartialResult import spark.storage.BlockManagerMaster import spark.storage.BlockManagerId +import util.{MetadataCleaner, TimeStampedHashMap} /** * A Scheduler subclass that implements stage-oriented scheduling. It computes a DAG of stages for @@ -22,7 +23,16 @@ import spark.storage.BlockManagerId * and to report fetch failures (the submitTasks method, and code to add CompletionEvents). */ private[spark] -class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with Logging { +class DAGScheduler( + taskSched: TaskScheduler, + mapOutputTracker: MapOutputTracker, + blockManagerMaster: BlockManagerMaster, + env: SparkEnv) + extends TaskSchedulerListener with Logging { + + def this(taskSched: TaskScheduler) { + this(taskSched, SparkEnv.get.mapOutputTracker, SparkEnv.get.blockManager.master, SparkEnv.get) + } taskSched.setListener(this) // Called by TaskScheduler to report task completions or failures. @@ -34,12 +44,12 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with eventQueue.put(CompletionEvent(task, reason, result, accumUpdates)) } - // Called by TaskScheduler when a host fails. - override def hostLost(host: String) { - eventQueue.put(HostLost(host)) + // Called by TaskScheduler when an executor fails. + override def executorLost(execId: String) { + eventQueue.put(ExecutorLost(execId)) } - // Called by TaskScheduler to cancel an entier TaskSet due to repeated failures. + // Called by TaskScheduler to cancel an entire TaskSet due to repeated failures. override def taskSetFailed(taskSet: TaskSet, reason: String) { eventQueue.put(TaskSetFailed(taskSet, reason)) } @@ -53,50 +63,59 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with // resubmit failed stages val POLL_TIMEOUT = 10L - private val lock = new Object // Used for access to the entire DAGScheduler - private val eventQueue = new LinkedBlockingQueue[DAGSchedulerEvent] val nextRunId = new AtomicInteger(0) val nextStageId = new AtomicInteger(0) - val idToStage = new HashMap[Int, Stage] + val idToStage = new TimeStampedHashMap[Int, Stage] - val shuffleToMapStage = new HashMap[Int, Stage] + val shuffleToMapStage = new TimeStampedHashMap[Int, Stage] var cacheLocs = new HashMap[Int, Array[List[String]]] - val env = SparkEnv.get - val cacheTracker = env.cacheTracker - val mapOutputTracker = env.mapOutputTracker - - val deadHosts = new HashSet[String] // TODO: The code currently assumes these can't come back; - // that's not going to be a realistic assumption in general + // For tracking failed nodes, we use the MapOutputTracker's generation number, which is + // sent with every task. When we detect a node failing, we note the current generation number + // and failed executor, increment it for new tasks, and use this to ignore stray ShuffleMapTask + // results. + // TODO: Garbage collect information about failure generations when we know there are no more + // stray messages to detect. + val failedGeneration = new HashMap[String, Long] val waiting = new HashSet[Stage] // Stages we need to run whose parents aren't done val running = new HashSet[Stage] // Stages we are running right now val failed = new HashSet[Stage] // Stages that must be resubmitted due to fetch failures - val pendingTasks = new HashMap[Stage, HashSet[Task[_]]] // Missing tasks from each stage + val pendingTasks = new TimeStampedHashMap[Stage, HashSet[Task[_]]] // Missing tasks from each stage var lastFetchFailureTime: Long = 0 // Used to wait a bit to avoid repeated resubmits val activeJobs = new HashSet[ActiveJob] val resultStageToJob = new HashMap[Stage, ActiveJob] + val metadataCleaner = new MetadataCleaner("DAGScheduler", this.cleanup) + // Start a thread to run the DAGScheduler event loop - new Thread("DAGScheduler") { - setDaemon(true) - override def run() { - DAGScheduler.this.run() - } - }.start() + def start() { + new Thread("DAGScheduler") { + setDaemon(true) + override def run() { + DAGScheduler.this.run() + } + }.start() + } - def getCacheLocs(rdd: RDD[_]): Array[List[String]] = { + private def getCacheLocs(rdd: RDD[_]): Array[List[String]] = { + if (!cacheLocs.contains(rdd.id)) { + val blockIds = rdd.splits.indices.map(index=> "rdd_%d_%d".format(rdd.id, index)).toArray + cacheLocs(rdd.id) = blockManagerMaster.getLocations(blockIds).map { + locations => locations.map(_.ip).toList + }.toArray + } cacheLocs(rdd.id) } - def updateCacheLocs() { - cacheLocs = cacheTracker.getLocationsSnapshot() + private def clearCacheLocs() { + cacheLocs.clear() } /** @@ -104,7 +123,7 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with * The priority value passed in will be used if the stage doesn't already exist with * a lower priority (we assume that priorities always increase across jobs for now). */ - def getShuffleMapStage(shuffleDep: ShuffleDependency[_,_], priority: Int): Stage = { + private def getShuffleMapStage(shuffleDep: ShuffleDependency[_,_], priority: Int): Stage = { shuffleToMapStage.get(shuffleDep.shuffleId) match { case Some(stage) => stage case None => @@ -119,12 +138,11 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with * as a result stage for the final RDD used directly in an action. The stage will also be given * the provided priority. */ - def newStage(rdd: RDD[_], shuffleDep: Option[ShuffleDependency[_,_]], priority: Int): Stage = { - // Kind of ugly: need to register RDDs with the cache and map output tracker here - // since we can't do it in the RDD constructor because # of splits is unknown - logInfo("Registering RDD " + rdd.id + " (" + rdd.origin + ")") - cacheTracker.registerRDD(rdd.id, rdd.splits.size) + private def newStage(rdd: RDD[_], shuffleDep: Option[ShuffleDependency[_,_]], priority: Int): Stage = { if (shuffleDep != None) { + // Kind of ugly: need to register RDDs with the cache and map output tracker here + // since we can't do it in the RDD constructor because # of splits is unknown + logInfo("Registering RDD " + rdd.id + " (" + rdd.origin + ")") mapOutputTracker.registerShuffle(shuffleDep.get.shuffleId, rdd.splits.size) } val id = nextStageId.getAndIncrement() @@ -137,7 +155,7 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with * Get or create the list of parent stages for a given RDD. The stages will be assigned the * provided priority if they haven't already been created with a lower priority. */ - def getParentStages(rdd: RDD[_], priority: Int): List[Stage] = { + private def getParentStages(rdd: RDD[_], priority: Int): List[Stage] = { val parents = new HashSet[Stage] val visited = new HashSet[RDD[_]] def visit(r: RDD[_]) { @@ -145,8 +163,6 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with visited += r // Kind of ugly: need to register RDDs with the cache here since // we can't do it in its constructor because # of splits is unknown - logInfo("Registering parent RDD " + r.id + " (" + r.origin + ")") - cacheTracker.registerRDD(r.id, r.splits.size) for (dep <- r.dependencies) { dep match { case shufDep: ShuffleDependency[_,_] => @@ -161,25 +177,22 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with parents.toList } - def getMissingParentStages(stage: Stage): List[Stage] = { + private def getMissingParentStages(stage: Stage): List[Stage] = { val missing = new HashSet[Stage] val visited = new HashSet[RDD[_]] def visit(rdd: RDD[_]) { if (!visited(rdd)) { visited += rdd - val locs = getCacheLocs(rdd) - for (p <- 0 until rdd.splits.size) { - if (locs(p) == Nil) { - for (dep <- rdd.dependencies) { - dep match { - case shufDep: ShuffleDependency[_,_] => - val mapStage = getShuffleMapStage(shufDep, stage.priority) - if (!mapStage.isAvailable) { - missing += mapStage - } - case narrowDep: NarrowDependency[_] => - visit(narrowDep.rdd) - } + if (getCacheLocs(rdd).contains(Nil)) { + for (dep <- rdd.dependencies) { + dep match { + case shufDep: ShuffleDependency[_,_] => + val mapStage = getShuffleMapStage(shufDep, stage.priority) + if (!mapStage.isAvailable) { + missing += mapStage + } + case narrowDep: NarrowDependency[_] => + visit(narrowDep.rdd) } } } @@ -189,23 +202,45 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with missing.toList } + /** + * Returns (and does not submit) a JobSubmitted event suitable to run a given job, and a + * JobWaiter whose getResult() method will return the result of the job when it is complete. + * + * The job is assumed to have at least one partition; zero partition jobs should be handled + * without a JobSubmitted event. + */ + private[scheduler] def prepareJob[T, U: ClassManifest]( + finalRdd: RDD[T], + func: (TaskContext, Iterator[T]) => U, + partitions: Seq[Int], + callSite: String, + allowLocal: Boolean, + resultHandler: (Int, U) => Unit) + : (JobSubmitted, JobWaiter[U]) = + { + assert(partitions.size > 0) + val waiter = new JobWaiter(partitions.size, resultHandler) + val func2 = func.asInstanceOf[(TaskContext, Iterator[_]) => _] + val toSubmit = JobSubmitted(finalRdd, func2, partitions.toArray, allowLocal, callSite, waiter) + return (toSubmit, waiter) + } + def runJob[T, U: ClassManifest]( finalRdd: RDD[T], func: (TaskContext, Iterator[T]) => U, partitions: Seq[Int], callSite: String, - allowLocal: Boolean) - : Array[U] = + allowLocal: Boolean, + resultHandler: (Int, U) => Unit) { if (partitions.size == 0) { - return new Array[U](0) + return } - val waiter = new JobWaiter(partitions.size) - val func2 = func.asInstanceOf[(TaskContext, Iterator[_]) => _] - eventQueue.put(JobSubmitted(finalRdd, func2, partitions.toArray, allowLocal, callSite, waiter)) - waiter.getResult() match { - case JobSucceeded(results: Seq[_]) => - return results.asInstanceOf[Seq[U]].toArray + val (toSubmit, waiter) = prepareJob( + finalRdd, func, partitions, callSite, allowLocal, resultHandler) + eventQueue.put(toSubmit) + waiter.awaitResult() match { + case JobSucceeded => {} case JobFailed(exception: Exception) => logInfo("Failed to run " + callSite) throw exception @@ -224,90 +259,117 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with val func2 = func.asInstanceOf[(TaskContext, Iterator[_]) => _] val partitions = (0 until rdd.splits.size).toArray eventQueue.put(JobSubmitted(rdd, func2, partitions, false, callSite, listener)) - return listener.getResult() // Will throw an exception if the job fails + return listener.awaitResult() // Will throw an exception if the job fails + } + + /** + * Process one event retrieved from the event queue. + * Returns true if we should stop the event loop. + */ + private[scheduler] def processEvent(event: DAGSchedulerEvent): Boolean = { + event match { + case JobSubmitted(finalRDD, func, partitions, allowLocal, callSite, listener) => + val runId = nextRunId.getAndIncrement() + val finalStage = newStage(finalRDD, None, runId) + val job = new ActiveJob(runId, finalStage, func, partitions, callSite, listener) + clearCacheLocs() + logInfo("Got job " + job.runId + " (" + callSite + ") with " + partitions.length + + " output partitions (allowLocal=" + allowLocal + ")") + logInfo("Final stage: " + finalStage + " (" + finalStage.origin + ")") + logInfo("Parents of final stage: " + finalStage.parents) + logInfo("Missing parents: " + getMissingParentStages(finalStage)) + if (allowLocal && finalStage.parents.size == 0 && partitions.length == 1) { + // Compute very short actions like first() or take() with no parent stages locally. + runLocally(job) + } else { + activeJobs += job + resultStageToJob(finalStage) = job + submitStage(finalStage) + } + + case ExecutorLost(execId) => + handleExecutorLost(execId) + + case completion: CompletionEvent => + handleTaskCompletion(completion) + + case TaskSetFailed(taskSet, reason) => + abortStage(idToStage(taskSet.stageId), reason) + + case StopDAGScheduler => + // Cancel any active jobs + for (job <- activeJobs) { + val error = new SparkException("Job cancelled because SparkContext was shut down") + job.listener.jobFailed(error) + } + return true + } + return false } /** + * Resubmit any failed stages. Ordinarily called after a small amount of time has passed since + * the last fetch failure. + */ + private[scheduler] def resubmitFailedStages() { + logInfo("Resubmitting failed stages") + clearCacheLocs() + val failed2 = failed.toArray + failed.clear() + for (stage <- failed2.sortBy(_.priority)) { + submitStage(stage) + } + } + + /** + * Check for waiting or failed stages which are now eligible for resubmission. + * Ordinarily run on every iteration of the event loop. + */ + private[scheduler] def submitWaitingStages() { + // TODO: We might want to run this less often, when we are sure that something has become + // runnable that wasn't before. + logTrace("Checking for newly runnable parent stages") + logTrace("running: " + running) + logTrace("waiting: " + waiting) + logTrace("failed: " + failed) + val waiting2 = waiting.toArray + waiting.clear() + for (stage <- waiting2.sortBy(_.priority)) { + submitStage(stage) + } + } + + + /** * The main event loop of the DAG scheduler, which waits for new-job / task-finished / failure * events and responds by launching tasks. This runs in a dedicated thread and receives events * via the eventQueue. */ - def run() { + private def run() { SparkEnv.set(env) while (true) { val event = eventQueue.poll(POLL_TIMEOUT, TimeUnit.MILLISECONDS) - val time = System.currentTimeMillis() // TODO: use a pluggable clock for testability if (event != null) { logDebug("Got event of type " + event.getClass.getName) } - event match { - case JobSubmitted(finalRDD, func, partitions, allowLocal, callSite, listener) => - val runId = nextRunId.getAndIncrement() - val finalStage = newStage(finalRDD, None, runId) - val job = new ActiveJob(runId, finalStage, func, partitions, callSite, listener) - updateCacheLocs() - logInfo("Got job " + job.runId + " (" + callSite + ") with " + partitions.length + - " output partitions") - logInfo("Final stage: " + finalStage + " (" + finalStage.origin + ")") - logInfo("Parents of final stage: " + finalStage.parents) - logInfo("Missing parents: " + getMissingParentStages(finalStage)) - if (allowLocal && finalStage.parents.size == 0 && partitions.length == 1) { - // Compute very short actions like first() or take() with no parent stages locally. - runLocally(job) - } else { - activeJobs += job - resultStageToJob(finalStage) = job - submitStage(finalStage) - } - - case HostLost(host) => - handleHostLost(host) - - case completion: CompletionEvent => - handleTaskCompletion(completion) - - case TaskSetFailed(taskSet, reason) => - abortStage(idToStage(taskSet.stageId), reason) - - case StopDAGScheduler => - // Cancel any active jobs - for (job <- activeJobs) { - val error = new SparkException("Job cancelled because SparkContext was shut down") - job.listener.jobFailed(error) - } + if (event != null) { + if (processEvent(event)) { return - - case null => - // queue.poll() timed out, ignore it + } } + val time = System.currentTimeMillis() // TODO: use a pluggable clock for testability // Periodically resubmit failed stages if some map output fetches have failed and we have // waited at least RESUBMIT_TIMEOUT. We wait for this short time because when a node fails, // tasks on many other nodes are bound to get a fetch failure, and they won't all get it at // the same time, so we want to make sure we've identified all the reduce tasks that depend // on the failed node. if (failed.size > 0 && time > lastFetchFailureTime + RESUBMIT_TIMEOUT) { - logInfo("Resubmitting failed stages") - updateCacheLocs() - val failed2 = failed.toArray - failed.clear() - for (stage <- failed2.sortBy(_.priority)) { - submitStage(stage) - } + resubmitFailedStages() } else { - // TODO: We might want to run this less often, when we are sure that something has become - // runnable that wasn't before. - logDebug("Checking for newly runnable parent stages") - logDebug("running: " + running) - logDebug("waiting: " + waiting) - logDebug("failed: " + failed) - val waiting2 = waiting.toArray - waiting.clear() - for (stage <- waiting2.sortBy(_.priority)) { - submitStage(stage) - } + submitWaitingStages() } } } @@ -317,7 +379,7 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with * We run the operation in a separate thread just in case it takes a bunch of time, so that we * don't block the DAGScheduler event loop or other concurrent jobs. */ - def runLocally(job: ActiveJob) { + private def runLocally(job: ActiveJob) { logInfo("Computing the requested partition locally") new Thread("Local computation of job " + job.runId) { override def run() { @@ -326,9 +388,12 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with val rdd = job.finalStage.rdd val split = rdd.splits(job.partitions(0)) val taskContext = new TaskContext(job.finalStage.id, job.partitions(0), 0) - val result = job.func(taskContext, rdd.iterator(split, taskContext)) - taskContext.executeOnCompleteCallbacks() - job.listener.taskSucceeded(0, result) + try { + val result = job.func(taskContext, rdd.iterator(split, taskContext)) + job.listener.taskSucceeded(0, result) + } finally { + taskContext.executeOnCompleteCallbacks() + } } catch { case e: Exception => job.listener.jobFailed(e) @@ -337,13 +402,14 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with }.start() } - def submitStage(stage: Stage) { + /** Submits stage, but first recursively submits any missing parents. */ + private def submitStage(stage: Stage) { logDebug("submitStage(" + stage + ")") if (!waiting(stage) && !running(stage) && !failed(stage)) { val missing = getMissingParentStages(stage).sortBy(_.id) logDebug("missing: " + missing) if (missing == Nil) { - logInfo("Submitting " + stage + " (" + stage.origin + "), which has no missing parents") + logInfo("Submitting " + stage + " (" + stage.rdd + "), which has no missing parents") submitMissingTasks(stage) running += stage } else { @@ -355,7 +421,8 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with } } - def submitMissingTasks(stage: Stage) { + /** Called when stage's parents are available and we can now do its task. */ + private def submitMissingTasks(stage: Stage) { logDebug("submitMissingTasks(" + stage + ")") // Get our pending tasks and remember them in our pendingTasks entry val myPending = pendingTasks.getOrElseUpdate(stage, new HashSet) @@ -376,11 +443,14 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with } } if (tasks.size > 0) { - logInfo("Submitting " + tasks.size + " missing tasks from " + stage) + logInfo("Submitting " + tasks.size + " missing tasks from " + stage + " (" + stage.rdd + ")") myPending ++= tasks logDebug("New pending tasks: " + myPending) taskSched.submitTasks( new TaskSet(tasks.toArray, stage.id, stage.newAttemptId(), stage.priority)) + if (!stage.submissionTime.isDefined) { + stage.submissionTime = Some(System.currentTimeMillis()) + } } else { logDebug("Stage " + stage + " is actually done; %b %d %d".format( stage.isAvailable, stage.numAvailableOutputs, stage.numPartitions)) @@ -392,9 +462,18 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with * Responds to a task finishing. This is called inside the event loop so it assumes that it can * modify the scheduler's internal state. Use taskEnded() to post a task end event from outside. */ - def handleTaskCompletion(event: CompletionEvent) { + private def handleTaskCompletion(event: CompletionEvent) { val task = event.task val stage = idToStage(task.stageId) + + def markStageAsFinished(stage: Stage) = { + val serviceTime = stage.submissionTime match { + case Some(t) => "%.03f".format((System.currentTimeMillis() - t) / 1000.0) + case _ => "Unkown" + } + logInfo("%s (%s) finished in %s s".format(stage, stage.origin, serviceTime)) + running -= stage + } event.reason match { case Success => logInfo("Completed " + task) @@ -409,13 +488,13 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with if (!job.finished(rt.outputId)) { job.finished(rt.outputId) = true job.numFinished += 1 - job.listener.taskSucceeded(rt.outputId, event.result) // If the whole job has finished, remove it if (job.numFinished == job.numPartitions) { activeJobs -= job resultStageToJob -= stage - running -= stage + markStageAsFinished(stage) } + job.listener.taskSucceeded(rt.outputId, event.result) } case None => logInfo("Ignoring result from " + rt + " because its job has finished") @@ -424,23 +503,32 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with case smt: ShuffleMapTask => val stage = idToStage(smt.stageId) val status = event.result.asInstanceOf[MapStatus] - val host = status.address.ip - logInfo("ShuffleMapTask finished with host " + host) - if (!deadHosts.contains(host)) { // TODO: Make sure hostnames are consistent with Mesos + val execId = status.location.executorId + logDebug("ShuffleMapTask finished on " + execId) + if (failedGeneration.contains(execId) && smt.generation <= failedGeneration(execId)) { + logInfo("Ignoring possibly bogus ShuffleMapTask completion from " + execId) + } else { stage.addOutputLoc(smt.partition, status) } if (running.contains(stage) && pendingTasks(stage).isEmpty) { - logInfo(stage + " (" + stage.origin + ") finished; looking for newly runnable stages") - running -= stage + markStageAsFinished(stage) + logInfo("looking for newly runnable stages") logInfo("running: " + running) logInfo("waiting: " + waiting) logInfo("failed: " + failed) if (stage.shuffleDep != None) { + // We supply true to increment the generation number here in case this is a + // recomputation of the map outputs. In that case, some nodes may have cached + // locations with holes (from when we detected the error) and will need the + // generation incremented to refetch them. + // TODO: Only increment the generation number if this is not the first time + // we registered these map outputs. mapOutputTracker.registerMapOutputs( stage.shuffleDep.get.shuffleId, - stage.outputLocs.map(list => if (list.isEmpty) null else list.head).toArray) + stage.outputLocs.map(list => if (list.isEmpty) null else list.head).toArray, + true) } - updateCacheLocs() + clearCacheLocs() if (stage.outputLocs.count(_ == Nil) != 0) { // Some tasks had failed; let's resubmit this stage // TODO: Lower-level scheduler should also deal with this @@ -459,7 +547,7 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with waiting --= newlyRunnable running ++= newlyRunnable for (stage <- newlyRunnable.sortBy(_.id)) { - logInfo("Submitting " + stage + " (" + stage.origin + "), which is now runnable") + logInfo("Submitting " + stage + " (" + stage.rdd + "), which is now runnable") submitMissingTasks(stage) } } @@ -490,9 +578,9 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with // Remember that a fetch failed now; this is used to resubmit the broken // stages later, after a small wait (to give other tasks the chance to fail) lastFetchFailureTime = System.currentTimeMillis() // TODO: Use pluggable clock - // TODO: mark the host as failed only if there were lots of fetch failures on it + // TODO: mark the executor as failed only if there were lots of fetch failures on it if (bmAddress != null) { - handleHostLost(bmAddress.ip) + handleExecutorLost(bmAddress.executorId, Some(task.generation)) } case other => @@ -502,22 +590,31 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with } /** - * Responds to a host being lost. This is called inside the event loop so it assumes that it can - * modify the scheduler's internal state. Use hostLost() to post a host lost event from outside. + * Responds to an executor being lost. This is called inside the event loop, so it assumes it can + * modify the scheduler's internal state. Use executorLost() to post a loss event from outside. + * + * Optionally the generation during which the failure was caught can be passed to avoid allowing + * stray fetch failures from possibly retriggering the detection of a node as lost. */ - def handleHostLost(host: String) { - if (!deadHosts.contains(host)) { - logInfo("Host lost: " + host) - deadHosts += host - env.blockManager.master.notifyADeadHost(host) + private def handleExecutorLost(execId: String, maybeGeneration: Option[Long] = None) { + val currentGeneration = maybeGeneration.getOrElse(mapOutputTracker.getGeneration) + if (!failedGeneration.contains(execId) || failedGeneration(execId) < currentGeneration) { + failedGeneration(execId) = currentGeneration + logInfo("Executor lost: %s (generation %d)".format(execId, currentGeneration)) + blockManagerMaster.removeExecutor(execId) // TODO: This will be really slow if we keep accumulating shuffle map stages for ((shuffleId, stage) <- shuffleToMapStage) { - stage.removeOutputsOnHost(host) + stage.removeOutputsOnExecutor(execId) val locs = stage.outputLocs.map(list => if (list.isEmpty) null else list.head).toArray mapOutputTracker.registerMapOutputs(shuffleId, locs, true) } - cacheTracker.cacheLost(host) - updateCacheLocs() + if (shuffleToMapStage.isEmpty) { + mapOutputTracker.incrementGeneration() + } + clearCacheLocs() + } else { + logDebug("Additional executor lost message for " + execId + + "(generation " + currentGeneration + ")") } } @@ -525,7 +622,7 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with * Aborts all jobs depending on a particular Stage. This is called in response to a task set * being cancelled by the TaskScheduler. Use taskSetFailed() to inject this event from outside. */ - def abortStage(failedStage: Stage, reason: String) { + private def abortStage(failedStage: Stage, reason: String) { val dependentStages = resultStageToJob.keys.filter(x => stageDependsOn(x, failedStage)).toSeq for (resultStage <- dependentStages) { val job = resultStageToJob(resultStage) @@ -541,7 +638,7 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with /** * Return true if one of stage's ancestors is target. */ - def stageDependsOn(stage: Stage, target: Stage): Boolean = { + private def stageDependsOn(stage: Stage, target: Stage): Boolean = { if (stage == target) { return true } @@ -568,7 +665,7 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with visitedRdds.contains(target.rdd) } - def getPreferredLocs(rdd: RDD[_], partition: Int): List[String] = { + private def getPreferredLocs(rdd: RDD[_], partition: Int): List[String] = { // If the partition is cached, return the cache locations val cached = getCacheLocs(rdd)(partition) if (cached != Nil) { @@ -594,8 +691,23 @@ class DAGScheduler(taskSched: TaskScheduler) extends TaskSchedulerListener with return Nil } + private def cleanup(cleanupTime: Long) { + var sizeBefore = idToStage.size + idToStage.clearOldValues(cleanupTime) + logInfo("idToStage " + sizeBefore + " --> " + idToStage.size) + + sizeBefore = shuffleToMapStage.size + shuffleToMapStage.clearOldValues(cleanupTime) + logInfo("shuffleToMapStage " + sizeBefore + " --> " + shuffleToMapStage.size) + + sizeBefore = pendingTasks.size + pendingTasks.clearOldValues(cleanupTime) + logInfo("pendingTasks " + sizeBefore + " --> " + pendingTasks.size) + } + def stop() { eventQueue.put(StopDAGScheduler) + metadataCleaner.cancel() taskSched.stop() } } |